1 /* 2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compilerDirectives.hpp" 32 #include "compiler/oopMap.hpp" 33 #include "memory/allocation.inline.hpp" 34 #include "opto/ad.hpp" 35 #include "opto/callnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/optoreg.hpp" 40 #include "opto/output.hpp" 41 #include "opto/regalloc.hpp" 42 #include "opto/runtime.hpp" 43 #include "opto/subnode.hpp" 44 #include "opto/type.hpp" 45 #include "runtime/handles.inline.hpp" 46 #include "utilities/xmlstream.hpp" 47 48 #ifndef PRODUCT 49 #define DEBUG_ARG(x) , x 50 #else 51 #define DEBUG_ARG(x) 52 #endif 53 54 // Convert Nodes to instruction bits and pass off to the VM 55 void Compile::Output() { 56 // RootNode goes 57 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 58 59 // The number of new nodes (mostly MachNop) is proportional to 60 // the number of java calls and inner loops which are aligned. 61 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 62 C->inner_loops()*(OptoLoopAlignment-1)), 63 "out of nodes before code generation" ) ) { 64 return; 65 } 66 // Make sure I can find the Start Node 67 Block *entry = _cfg->get_block(1); 68 Block *broot = _cfg->get_root_block(); 69 70 const StartNode *start = entry->head()->as_Start(); 71 72 // Replace StartNode with prolog 73 MachPrologNode *prolog = new MachPrologNode(); 74 entry->map_node(prolog, 0); 75 _cfg->map_node_to_block(prolog, entry); 76 _cfg->unmap_node_from_block(start); // start is no longer in any block 77 78 // Virtual methods need an unverified entry point 79 80 if( is_osr_compilation() ) { 81 if( PoisonOSREntry ) { 82 // TODO: Should use a ShouldNotReachHereNode... 83 _cfg->insert( broot, 0, new MachBreakpointNode() ); 84 } 85 } else { 86 if( _method && !_method->flags().is_static() ) { 87 // Insert unvalidated entry point 88 _cfg->insert( broot, 0, new MachUEPNode() ); 89 } 90 91 } 92 93 // Break before main entry point 94 if ((_method && C->directive()->BreakAtExecuteOption) || 95 (OptoBreakpoint && is_method_compilation()) || 96 (OptoBreakpointOSR && is_osr_compilation()) || 97 (OptoBreakpointC2R && !_method) ) { 98 // checking for _method means that OptoBreakpoint does not apply to 99 // runtime stubs or frame converters 100 _cfg->insert( entry, 1, new MachBreakpointNode() ); 101 } 102 103 // Insert epilogs before every return 104 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 105 Block* block = _cfg->get_block(i); 106 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 107 Node* m = block->end(); 108 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 109 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 110 block->add_inst(epilog); 111 _cfg->map_node_to_block(epilog, block); 112 } 113 } 114 } 115 116 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 117 blk_starts[0] = 0; 118 119 // Initialize code buffer and process short branches. 120 CodeBuffer* cb = init_buffer(blk_starts); 121 122 if (cb == NULL || failing()) { 123 return; 124 } 125 126 ScheduleAndBundle(); 127 128 #ifndef PRODUCT 129 if (trace_opto_output()) { 130 tty->print("\n---- After ScheduleAndBundle ----\n"); 131 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 132 tty->print("\nBB#%03d:\n", i); 133 Block* block = _cfg->get_block(i); 134 for (uint j = 0; j < block->number_of_nodes(); j++) { 135 Node* n = block->get_node(j); 136 OptoReg::Name reg = _regalloc->get_reg_first(n); 137 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 138 n->dump(); 139 } 140 } 141 } 142 #endif 143 144 if (failing()) { 145 return; 146 } 147 148 BuildOopMaps(); 149 150 if (failing()) { 151 return; 152 } 153 154 fill_buffer(cb, blk_starts); 155 } 156 157 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 158 // Determine if we need to generate a stack overflow check. 159 // Do it if the method is not a stub function and 160 // has java calls or has frame size > vm_page_size/8. 161 // The debug VM checks that deoptimization doesn't trigger an 162 // unexpected stack overflow (compiled method stack banging should 163 // guarantee it doesn't happen) so we always need the stack bang in 164 // a debug VM. 165 return (UseStackBanging && stub_function() == NULL && 166 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 167 DEBUG_ONLY(|| true))); 168 } 169 170 bool Compile::need_register_stack_bang() const { 171 // Determine if we need to generate a register stack overflow check. 172 // This is only used on architectures which have split register 173 // and memory stacks (ie. IA64). 174 // Bang if the method is not a stub function and has java calls 175 return (stub_function() == NULL && has_java_calls()); 176 } 177 178 179 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 180 // of a loop. When aligning a loop we need to provide enough instructions 181 // in cpu's fetch buffer to feed decoders. The loop alignment could be 182 // avoided if we have enough instructions in fetch buffer at the head of a loop. 183 // By default, the size is set to 999999 by Block's constructor so that 184 // a loop will be aligned if the size is not reset here. 185 // 186 // Note: Mach instructions could contain several HW instructions 187 // so the size is estimated only. 188 // 189 void Compile::compute_loop_first_inst_sizes() { 190 // The next condition is used to gate the loop alignment optimization. 191 // Don't aligned a loop if there are enough instructions at the head of a loop 192 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 193 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 194 // equal to 11 bytes which is the largest address NOP instruction. 195 if (MaxLoopPad < OptoLoopAlignment - 1) { 196 uint last_block = _cfg->number_of_blocks() - 1; 197 for (uint i = 1; i <= last_block; i++) { 198 Block* block = _cfg->get_block(i); 199 // Check the first loop's block which requires an alignment. 200 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 201 uint sum_size = 0; 202 uint inst_cnt = NumberOfLoopInstrToAlign; 203 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 204 205 // Check subsequent fallthrough blocks if the loop's first 206 // block(s) does not have enough instructions. 207 Block *nb = block; 208 while(inst_cnt > 0 && 209 i < last_block && 210 !_cfg->get_block(i + 1)->has_loop_alignment() && 211 !nb->has_successor(block)) { 212 i++; 213 nb = _cfg->get_block(i); 214 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 215 } // while( inst_cnt > 0 && i < last_block ) 216 217 block->set_first_inst_size(sum_size); 218 } // f( b->head()->is_Loop() ) 219 } // for( i <= last_block ) 220 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 221 } 222 223 // The architecture description provides short branch variants for some long 224 // branch instructions. Replace eligible long branches with short branches. 225 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 226 // Compute size of each block, method size, and relocation information size 227 uint nblocks = _cfg->number_of_blocks(); 228 229 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 230 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 231 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 232 233 // Collect worst case block paddings 234 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 235 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 236 237 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 238 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 239 240 bool has_short_branch_candidate = false; 241 242 // Initialize the sizes to 0 243 code_size = 0; // Size in bytes of generated code 244 stub_size = 0; // Size in bytes of all stub entries 245 // Size in bytes of all relocation entries, including those in local stubs. 246 // Start with 2-bytes of reloc info for the unvalidated entry point 247 reloc_size = 1; // Number of relocation entries 248 249 // Make three passes. The first computes pessimistic blk_starts, 250 // relative jmp_offset and reloc_size information. The second performs 251 // short branch substitution using the pessimistic sizing. The 252 // third inserts nops where needed. 253 254 // Step one, perform a pessimistic sizing pass. 255 uint last_call_adr = max_juint; 256 uint last_avoid_back_to_back_adr = max_juint; 257 uint nop_size = (new MachNopNode())->size(_regalloc); 258 for (uint i = 0; i < nblocks; i++) { // For all blocks 259 Block* block = _cfg->get_block(i); 260 261 // During short branch replacement, we store the relative (to blk_starts) 262 // offset of jump in jmp_offset, rather than the absolute offset of jump. 263 // This is so that we do not need to recompute sizes of all nodes when 264 // we compute correct blk_starts in our next sizing pass. 265 jmp_offset[i] = 0; 266 jmp_size[i] = 0; 267 jmp_nidx[i] = -1; 268 DEBUG_ONLY( jmp_target[i] = 0; ) 269 DEBUG_ONLY( jmp_rule[i] = 0; ) 270 271 // Sum all instruction sizes to compute block size 272 uint last_inst = block->number_of_nodes(); 273 uint blk_size = 0; 274 for (uint j = 0; j < last_inst; j++) { 275 Node* nj = block->get_node(j); 276 // Handle machine instruction nodes 277 if (nj->is_Mach()) { 278 MachNode *mach = nj->as_Mach(); 279 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 280 reloc_size += mach->reloc(); 281 if (mach->is_MachCall()) { 282 // add size information for trampoline stub 283 // class CallStubImpl is platform-specific and defined in the *.ad files. 284 stub_size += CallStubImpl::size_call_trampoline(); 285 reloc_size += CallStubImpl::reloc_call_trampoline(); 286 287 MachCallNode *mcall = mach->as_MachCall(); 288 // This destination address is NOT PC-relative 289 290 mcall->method_set((intptr_t)mcall->entry_point()); 291 292 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 293 stub_size += CompiledStaticCall::to_interp_stub_size(); 294 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 295 #if INCLUDE_AOT 296 stub_size += CompiledStaticCall::to_aot_stub_size(); 297 reloc_size += CompiledStaticCall::reloc_to_aot_stub(); 298 #endif 299 } 300 } else if (mach->is_MachSafePoint()) { 301 // If call/safepoint are adjacent, account for possible 302 // nop to disambiguate the two safepoints. 303 // ScheduleAndBundle() can rearrange nodes in a block, 304 // check for all offsets inside this block. 305 if (last_call_adr >= blk_starts[i]) { 306 blk_size += nop_size; 307 } 308 } 309 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 310 // Nop is inserted between "avoid back to back" instructions. 311 // ScheduleAndBundle() can rearrange nodes in a block, 312 // check for all offsets inside this block. 313 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 314 blk_size += nop_size; 315 } 316 } 317 if (mach->may_be_short_branch()) { 318 if (!nj->is_MachBranch()) { 319 #ifndef PRODUCT 320 nj->dump(3); 321 #endif 322 Unimplemented(); 323 } 324 assert(jmp_nidx[i] == -1, "block should have only one branch"); 325 jmp_offset[i] = blk_size; 326 jmp_size[i] = nj->size(_regalloc); 327 jmp_nidx[i] = j; 328 has_short_branch_candidate = true; 329 } 330 } 331 blk_size += nj->size(_regalloc); 332 // Remember end of call offset 333 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 334 last_call_adr = blk_starts[i]+blk_size; 335 } 336 // Remember end of avoid_back_to_back offset 337 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 338 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 339 } 340 } 341 342 // When the next block starts a loop, we may insert pad NOP 343 // instructions. Since we cannot know our future alignment, 344 // assume the worst. 345 if (i < nblocks - 1) { 346 Block* nb = _cfg->get_block(i + 1); 347 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 348 if (max_loop_pad > 0) { 349 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 350 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 351 // If either is the last instruction in this block, bump by 352 // max_loop_pad in lock-step with blk_size, so sizing 353 // calculations in subsequent blocks still can conservatively 354 // detect that it may the last instruction in this block. 355 if (last_call_adr == blk_starts[i]+blk_size) { 356 last_call_adr += max_loop_pad; 357 } 358 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 359 last_avoid_back_to_back_adr += max_loop_pad; 360 } 361 blk_size += max_loop_pad; 362 block_worst_case_pad[i + 1] = max_loop_pad; 363 } 364 } 365 366 // Save block size; update total method size 367 blk_starts[i+1] = blk_starts[i]+blk_size; 368 } 369 370 // Step two, replace eligible long jumps. 371 bool progress = true; 372 uint last_may_be_short_branch_adr = max_juint; 373 while (has_short_branch_candidate && progress) { 374 progress = false; 375 has_short_branch_candidate = false; 376 int adjust_block_start = 0; 377 for (uint i = 0; i < nblocks; i++) { 378 Block* block = _cfg->get_block(i); 379 int idx = jmp_nidx[i]; 380 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 381 if (mach != NULL && mach->may_be_short_branch()) { 382 #ifdef ASSERT 383 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 384 int j; 385 // Find the branch; ignore trailing NOPs. 386 for (j = block->number_of_nodes()-1; j>=0; j--) { 387 Node* n = block->get_node(j); 388 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 389 break; 390 } 391 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 392 #endif 393 int br_size = jmp_size[i]; 394 int br_offs = blk_starts[i] + jmp_offset[i]; 395 396 // This requires the TRUE branch target be in succs[0] 397 uint bnum = block->non_connector_successor(0)->_pre_order; 398 int offset = blk_starts[bnum] - br_offs; 399 if (bnum > i) { // adjust following block's offset 400 offset -= adjust_block_start; 401 } 402 403 // This block can be a loop header, account for the padding 404 // in the previous block. 405 int block_padding = block_worst_case_pad[i]; 406 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 407 // In the following code a nop could be inserted before 408 // the branch which will increase the backward distance. 409 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 410 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 411 412 if (needs_padding && offset <= 0) 413 offset -= nop_size; 414 415 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 416 // We've got a winner. Replace this branch. 417 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 418 419 // Update the jmp_size. 420 int new_size = replacement->size(_regalloc); 421 int diff = br_size - new_size; 422 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 423 // Conservatively take into account padding between 424 // avoid_back_to_back branches. Previous branch could be 425 // converted into avoid_back_to_back branch during next 426 // rounds. 427 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 428 jmp_offset[i] += nop_size; 429 diff -= nop_size; 430 } 431 adjust_block_start += diff; 432 block->map_node(replacement, idx); 433 mach->subsume_by(replacement, C); 434 mach = replacement; 435 progress = true; 436 437 jmp_size[i] = new_size; 438 DEBUG_ONLY( jmp_target[i] = bnum; ); 439 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 440 } else { 441 // The jump distance is not short, try again during next iteration. 442 has_short_branch_candidate = true; 443 } 444 } // (mach->may_be_short_branch()) 445 if (mach != NULL && (mach->may_be_short_branch() || 446 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 447 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 448 } 449 blk_starts[i+1] -= adjust_block_start; 450 } 451 } 452 453 #ifdef ASSERT 454 for (uint i = 0; i < nblocks; i++) { // For all blocks 455 if (jmp_target[i] != 0) { 456 int br_size = jmp_size[i]; 457 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 458 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 459 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 460 } 461 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 462 } 463 } 464 #endif 465 466 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 467 // after ScheduleAndBundle(). 468 469 // ------------------ 470 // Compute size for code buffer 471 code_size = blk_starts[nblocks]; 472 473 // Relocation records 474 reloc_size += 1; // Relo entry for exception handler 475 476 // Adjust reloc_size to number of record of relocation info 477 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 478 // a relocation index. 479 // The CodeBuffer will expand the locs array if this estimate is too low. 480 reloc_size *= 10 / sizeof(relocInfo); 481 } 482 483 //------------------------------FillLocArray----------------------------------- 484 // Create a bit of debug info and append it to the array. The mapping is from 485 // Java local or expression stack to constant, register or stack-slot. For 486 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 487 // entry has been taken care of and caller should skip it). 488 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 489 // This should never have accepted Bad before 490 assert(OptoReg::is_valid(regnum), "location must be valid"); 491 return (OptoReg::is_reg(regnum)) 492 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 493 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 494 } 495 496 497 ObjectValue* 498 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 499 for (int i = 0; i < objs->length(); i++) { 500 assert(objs->at(i)->is_object(), "corrupt object cache"); 501 ObjectValue* sv = (ObjectValue*) objs->at(i); 502 if (sv->id() == id) { 503 return sv; 504 } 505 } 506 // Otherwise.. 507 return NULL; 508 } 509 510 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 511 ObjectValue* sv ) { 512 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 513 objs->append(sv); 514 } 515 516 517 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 518 GrowableArray<ScopeValue*> *array, 519 GrowableArray<ScopeValue*> *objs ) { 520 assert( local, "use _top instead of null" ); 521 if (array->length() != idx) { 522 assert(array->length() == idx + 1, "Unexpected array count"); 523 // Old functionality: 524 // return 525 // New functionality: 526 // Assert if the local is not top. In product mode let the new node 527 // override the old entry. 528 assert(local == top(), "LocArray collision"); 529 if (local == top()) { 530 return; 531 } 532 array->pop(); 533 } 534 const Type *t = local->bottom_type(); 535 536 // Is it a safepoint scalar object node? 537 if (local->is_SafePointScalarObject()) { 538 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 539 540 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 541 if (sv == NULL) { 542 ciKlass* cik = t->is_oopptr()->klass(); 543 assert(cik->is_instance_klass() || 544 cik->is_array_klass(), "Not supported allocation."); 545 sv = new ObjectValue(spobj->_idx, 546 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 547 Compile::set_sv_for_object_node(objs, sv); 548 549 uint first_ind = spobj->first_index(sfpt->jvms()); 550 for (uint i = 0; i < spobj->n_fields(); i++) { 551 Node* fld_node = sfpt->in(first_ind+i); 552 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 553 } 554 } 555 array->append(sv); 556 return; 557 } 558 559 // Grab the register number for the local 560 OptoReg::Name regnum = _regalloc->get_reg_first(local); 561 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 562 // Record the double as two float registers. 563 // The register mask for such a value always specifies two adjacent 564 // float registers, with the lower register number even. 565 // Normally, the allocation of high and low words to these registers 566 // is irrelevant, because nearly all operations on register pairs 567 // (e.g., StoreD) treat them as a single unit. 568 // Here, we assume in addition that the words in these two registers 569 // stored "naturally" (by operations like StoreD and double stores 570 // within the interpreter) such that the lower-numbered register 571 // is written to the lower memory address. This may seem like 572 // a machine dependency, but it is not--it is a requirement on 573 // the author of the <arch>.ad file to ensure that, for every 574 // even/odd double-register pair to which a double may be allocated, 575 // the word in the even single-register is stored to the first 576 // memory word. (Note that register numbers are completely 577 // arbitrary, and are not tied to any machine-level encodings.) 578 #ifdef _LP64 579 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 580 array->append(new ConstantIntValue(0)); 581 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 582 } else if ( t->base() == Type::Long ) { 583 array->append(new ConstantIntValue(0)); 584 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 585 } else if ( t->base() == Type::RawPtr ) { 586 // jsr/ret return address which must be restored into a the full 587 // width 64-bit stack slot. 588 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 589 } 590 #else //_LP64 591 #ifdef SPARC 592 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 593 // For SPARC we have to swap high and low words for 594 // long values stored in a single-register (g0-g7). 595 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 596 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 597 } else 598 #endif //SPARC 599 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 600 // Repack the double/long as two jints. 601 // The convention the interpreter uses is that the second local 602 // holds the first raw word of the native double representation. 603 // This is actually reasonable, since locals and stack arrays 604 // grow downwards in all implementations. 605 // (If, on some machine, the interpreter's Java locals or stack 606 // were to grow upwards, the embedded doubles would be word-swapped.) 607 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 608 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 609 } 610 #endif //_LP64 611 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 612 OptoReg::is_reg(regnum) ) { 613 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 614 ? Location::float_in_dbl : Location::normal )); 615 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 616 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 617 ? Location::int_in_long : Location::normal )); 618 } else if( t->base() == Type::NarrowOop ) { 619 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 620 } else { 621 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 622 } 623 return; 624 } 625 626 // No register. It must be constant data. 627 switch (t->base()) { 628 case Type::Half: // Second half of a double 629 ShouldNotReachHere(); // Caller should skip 2nd halves 630 break; 631 case Type::AnyPtr: 632 array->append(new ConstantOopWriteValue(NULL)); 633 break; 634 case Type::AryPtr: 635 case Type::ValueTypePtr: 636 case Type::InstPtr: // fall through 637 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 638 break; 639 case Type::NarrowOop: 640 if (t == TypeNarrowOop::NULL_PTR) { 641 array->append(new ConstantOopWriteValue(NULL)); 642 } else { 643 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 644 } 645 break; 646 case Type::Int: 647 array->append(new ConstantIntValue(t->is_int()->get_con())); 648 break; 649 case Type::RawPtr: 650 // A return address (T_ADDRESS). 651 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 652 #ifdef _LP64 653 // Must be restored to the full-width 64-bit stack slot. 654 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 655 #else 656 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 657 #endif 658 break; 659 case Type::FloatCon: { 660 float f = t->is_float_constant()->getf(); 661 array->append(new ConstantIntValue(jint_cast(f))); 662 break; 663 } 664 case Type::DoubleCon: { 665 jdouble d = t->is_double_constant()->getd(); 666 #ifdef _LP64 667 array->append(new ConstantIntValue(0)); 668 array->append(new ConstantDoubleValue(d)); 669 #else 670 // Repack the double as two jints. 671 // The convention the interpreter uses is that the second local 672 // holds the first raw word of the native double representation. 673 // This is actually reasonable, since locals and stack arrays 674 // grow downwards in all implementations. 675 // (If, on some machine, the interpreter's Java locals or stack 676 // were to grow upwards, the embedded doubles would be word-swapped.) 677 jlong_accessor acc; 678 acc.long_value = jlong_cast(d); 679 array->append(new ConstantIntValue(acc.words[1])); 680 array->append(new ConstantIntValue(acc.words[0])); 681 #endif 682 break; 683 } 684 case Type::Long: { 685 jlong d = t->is_long()->get_con(); 686 #ifdef _LP64 687 array->append(new ConstantIntValue(0)); 688 array->append(new ConstantLongValue(d)); 689 #else 690 // Repack the long as two jints. 691 // The convention the interpreter uses is that the second local 692 // holds the first raw word of the native double representation. 693 // This is actually reasonable, since locals and stack arrays 694 // grow downwards in all implementations. 695 // (If, on some machine, the interpreter's Java locals or stack 696 // were to grow upwards, the embedded doubles would be word-swapped.) 697 jlong_accessor acc; 698 acc.long_value = d; 699 array->append(new ConstantIntValue(acc.words[1])); 700 array->append(new ConstantIntValue(acc.words[0])); 701 #endif 702 break; 703 } 704 case Type::Top: // Add an illegal value here 705 array->append(new LocationValue(Location())); 706 break; 707 default: 708 ShouldNotReachHere(); 709 break; 710 } 711 } 712 713 // Determine if this node starts a bundle 714 bool Compile::starts_bundle(const Node *n) const { 715 return (_node_bundling_limit > n->_idx && 716 _node_bundling_base[n->_idx].starts_bundle()); 717 } 718 719 //--------------------------Process_OopMap_Node-------------------------------- 720 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 721 722 // Handle special safepoint nodes for synchronization 723 MachSafePointNode *sfn = mach->as_MachSafePoint(); 724 MachCallNode *mcall; 725 726 int safepoint_pc_offset = current_offset; 727 bool is_method_handle_invoke = false; 728 bool return_oop = false; 729 730 // Add the safepoint in the DebugInfoRecorder 731 if( !mach->is_MachCall() ) { 732 mcall = NULL; 733 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 734 } else { 735 mcall = mach->as_MachCall(); 736 737 // Is the call a MethodHandle call? 738 if (mcall->is_MachCallJava()) { 739 if (mcall->as_MachCallJava()->_method_handle_invoke) { 740 assert(has_method_handle_invokes(), "must have been set during call generation"); 741 is_method_handle_invoke = true; 742 } 743 } 744 745 // Check if a call returns an object. 746 if (mcall->returns_pointer()) { 747 return_oop = true; 748 } 749 safepoint_pc_offset += mcall->ret_addr_offset(); 750 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 751 } 752 753 // Loop over the JVMState list to add scope information 754 // Do not skip safepoints with a NULL method, they need monitor info 755 JVMState* youngest_jvms = sfn->jvms(); 756 int max_depth = youngest_jvms->depth(); 757 758 // Allocate the object pool for scalar-replaced objects -- the map from 759 // small-integer keys (which can be recorded in the local and ostack 760 // arrays) to descriptions of the object state. 761 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 762 763 // Visit scopes from oldest to youngest. 764 for (int depth = 1; depth <= max_depth; depth++) { 765 JVMState* jvms = youngest_jvms->of_depth(depth); 766 int idx; 767 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 768 // Safepoints that do not have method() set only provide oop-map and monitor info 769 // to support GC; these do not support deoptimization. 770 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 771 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 772 int num_mon = jvms->nof_monitors(); 773 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 774 "JVMS local count must match that of the method"); 775 776 // Add Local and Expression Stack Information 777 778 // Insert locals into the locarray 779 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 780 for( idx = 0; idx < num_locs; idx++ ) { 781 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 782 } 783 784 // Insert expression stack entries into the exparray 785 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 786 for( idx = 0; idx < num_exps; idx++ ) { 787 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 788 } 789 790 // Add in mappings of the monitors 791 assert( !method || 792 !method->is_synchronized() || 793 method->is_native() || 794 num_mon > 0 || 795 !GenerateSynchronizationCode, 796 "monitors must always exist for synchronized methods"); 797 798 // Build the growable array of ScopeValues for exp stack 799 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 800 801 // Loop over monitors and insert into array 802 for (idx = 0; idx < num_mon; idx++) { 803 // Grab the node that defines this monitor 804 Node* box_node = sfn->monitor_box(jvms, idx); 805 Node* obj_node = sfn->monitor_obj(jvms, idx); 806 807 // Create ScopeValue for object 808 ScopeValue *scval = NULL; 809 810 if (obj_node->is_SafePointScalarObject()) { 811 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 812 scval = Compile::sv_for_node_id(objs, spobj->_idx); 813 if (scval == NULL) { 814 const Type *t = spobj->bottom_type(); 815 ciKlass* cik = t->is_oopptr()->klass(); 816 assert(cik->is_instance_klass() || 817 cik->is_array_klass(), "Not supported allocation."); 818 ObjectValue* sv = new ObjectValue(spobj->_idx, 819 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 820 Compile::set_sv_for_object_node(objs, sv); 821 822 uint first_ind = spobj->first_index(youngest_jvms); 823 for (uint i = 0; i < spobj->n_fields(); i++) { 824 Node* fld_node = sfn->in(first_ind+i); 825 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 826 } 827 scval = sv; 828 } 829 } else if (!obj_node->is_Con()) { 830 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 831 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 832 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 833 } else { 834 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 835 } 836 } else { 837 const TypePtr *tp = obj_node->get_ptr_type(); 838 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 839 } 840 841 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 842 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 843 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 844 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 845 } 846 847 // We dump the object pool first, since deoptimization reads it in first. 848 debug_info()->dump_object_pool(objs); 849 850 // Build first class objects to pass to scope 851 DebugToken *locvals = debug_info()->create_scope_values(locarray); 852 DebugToken *expvals = debug_info()->create_scope_values(exparray); 853 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 854 855 // Make method available for all Safepoints 856 ciMethod* scope_method = method ? method : _method; 857 // Describe the scope here 858 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 859 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 860 // Now we can describe the scope. 861 methodHandle null_mh; 862 bool rethrow_exception = false; 863 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, locvals, expvals, monvals); 864 } // End jvms loop 865 866 // Mark the end of the scope set. 867 debug_info()->end_safepoint(safepoint_pc_offset); 868 } 869 870 871 872 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 873 class NonSafepointEmitter { 874 Compile* C; 875 JVMState* _pending_jvms; 876 int _pending_offset; 877 878 void emit_non_safepoint(); 879 880 public: 881 NonSafepointEmitter(Compile* compile) { 882 this->C = compile; 883 _pending_jvms = NULL; 884 _pending_offset = 0; 885 } 886 887 void observe_instruction(Node* n, int pc_offset) { 888 if (!C->debug_info()->recording_non_safepoints()) return; 889 890 Node_Notes* nn = C->node_notes_at(n->_idx); 891 if (nn == NULL || nn->jvms() == NULL) return; 892 if (_pending_jvms != NULL && 893 _pending_jvms->same_calls_as(nn->jvms())) { 894 // Repeated JVMS? Stretch it up here. 895 _pending_offset = pc_offset; 896 } else { 897 if (_pending_jvms != NULL && 898 _pending_offset < pc_offset) { 899 emit_non_safepoint(); 900 } 901 _pending_jvms = NULL; 902 if (pc_offset > C->debug_info()->last_pc_offset()) { 903 // This is the only way _pending_jvms can become non-NULL: 904 _pending_jvms = nn->jvms(); 905 _pending_offset = pc_offset; 906 } 907 } 908 } 909 910 // Stay out of the way of real safepoints: 911 void observe_safepoint(JVMState* jvms, int pc_offset) { 912 if (_pending_jvms != NULL && 913 !_pending_jvms->same_calls_as(jvms) && 914 _pending_offset < pc_offset) { 915 emit_non_safepoint(); 916 } 917 _pending_jvms = NULL; 918 } 919 920 void flush_at_end() { 921 if (_pending_jvms != NULL) { 922 emit_non_safepoint(); 923 } 924 _pending_jvms = NULL; 925 } 926 }; 927 928 void NonSafepointEmitter::emit_non_safepoint() { 929 JVMState* youngest_jvms = _pending_jvms; 930 int pc_offset = _pending_offset; 931 932 // Clear it now: 933 _pending_jvms = NULL; 934 935 DebugInformationRecorder* debug_info = C->debug_info(); 936 assert(debug_info->recording_non_safepoints(), "sanity"); 937 938 debug_info->add_non_safepoint(pc_offset); 939 int max_depth = youngest_jvms->depth(); 940 941 // Visit scopes from oldest to youngest. 942 for (int depth = 1; depth <= max_depth; depth++) { 943 JVMState* jvms = youngest_jvms->of_depth(depth); 944 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 945 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 946 methodHandle null_mh; 947 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 948 } 949 950 // Mark the end of the scope set. 951 debug_info->end_non_safepoint(pc_offset); 952 } 953 954 //------------------------------init_buffer------------------------------------ 955 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 956 957 // Set the initially allocated size 958 int code_req = initial_code_capacity; 959 int locs_req = initial_locs_capacity; 960 int stub_req = initial_stub_capacity; 961 int const_req = initial_const_capacity; 962 963 int pad_req = NativeCall::instruction_size; 964 // The extra spacing after the code is necessary on some platforms. 965 // Sometimes we need to patch in a jump after the last instruction, 966 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 967 968 // Compute the byte offset where we can store the deopt pc. 969 if (fixed_slots() != 0) { 970 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 971 } 972 973 // Compute prolog code size 974 _method_size = 0; 975 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 976 #if defined(IA64) && !defined(AIX) 977 if (save_argument_registers()) { 978 // 4815101: this is a stub with implicit and unknown precision fp args. 979 // The usual spill mechanism can only generate stfd's in this case, which 980 // doesn't work if the fp reg to spill contains a single-precision denorm. 981 // Instead, we hack around the normal spill mechanism using stfspill's and 982 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 983 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 984 // 985 // If we ever implement 16-byte 'registers' == stack slots, we can 986 // get rid of this hack and have SpillCopy generate stfspill/ldffill 987 // instead of stfd/stfs/ldfd/ldfs. 988 _frame_slots += 8*(16/BytesPerInt); 989 } 990 #endif 991 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 992 993 if (has_mach_constant_base_node()) { 994 uint add_size = 0; 995 // Fill the constant table. 996 // Note: This must happen before shorten_branches. 997 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 998 Block* b = _cfg->get_block(i); 999 1000 for (uint j = 0; j < b->number_of_nodes(); j++) { 1001 Node* n = b->get_node(j); 1002 1003 // If the node is a MachConstantNode evaluate the constant 1004 // value section. 1005 if (n->is_MachConstant()) { 1006 MachConstantNode* machcon = n->as_MachConstant(); 1007 machcon->eval_constant(C); 1008 } else if (n->is_Mach()) { 1009 // On Power there are more nodes that issue constants. 1010 add_size += (n->as_Mach()->ins_num_consts() * 8); 1011 } 1012 } 1013 } 1014 1015 // Calculate the offsets of the constants and the size of the 1016 // constant table (including the padding to the next section). 1017 constant_table().calculate_offsets_and_size(); 1018 const_req = constant_table().size() + add_size; 1019 } 1020 1021 // Initialize the space for the BufferBlob used to find and verify 1022 // instruction size in MachNode::emit_size() 1023 init_scratch_buffer_blob(const_req); 1024 if (failing()) return NULL; // Out of memory 1025 1026 // Pre-compute the length of blocks and replace 1027 // long branches with short if machine supports it. 1028 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1029 1030 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1031 // class HandlerImpl is platform-specific and defined in the *.ad files. 1032 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1033 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1034 stub_req += MAX_stubs_size; // ensure per-stub margin 1035 code_req += MAX_inst_size; // ensure per-instruction margin 1036 1037 if (StressCodeBuffers) 1038 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1039 1040 int total_req = 1041 const_req + 1042 code_req + 1043 pad_req + 1044 stub_req + 1045 exception_handler_req + 1046 deopt_handler_req; // deopt handler 1047 1048 if (has_method_handle_invokes()) 1049 total_req += deopt_handler_req; // deopt MH handler 1050 1051 CodeBuffer* cb = code_buffer(); 1052 cb->initialize(total_req, locs_req); 1053 1054 // Have we run out of code space? 1055 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1056 C->record_failure("CodeCache is full"); 1057 return NULL; 1058 } 1059 // Configure the code buffer. 1060 cb->initialize_consts_size(const_req); 1061 cb->initialize_stubs_size(stub_req); 1062 cb->initialize_oop_recorder(env()->oop_recorder()); 1063 1064 // fill in the nop array for bundling computations 1065 MachNode *_nop_list[Bundle::_nop_count]; 1066 Bundle::initialize_nops(_nop_list); 1067 1068 return cb; 1069 } 1070 1071 //------------------------------fill_buffer------------------------------------ 1072 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1073 // blk_starts[] contains offsets calculated during short branches processing, 1074 // offsets should not be increased during following steps. 1075 1076 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1077 // of a loop. It is used to determine the padding for loop alignment. 1078 compute_loop_first_inst_sizes(); 1079 1080 // Create oopmap set. 1081 _oop_map_set = new OopMapSet(); 1082 1083 // !!!!! This preserves old handling of oopmaps for now 1084 debug_info()->set_oopmaps(_oop_map_set); 1085 1086 uint nblocks = _cfg->number_of_blocks(); 1087 // Count and start of implicit null check instructions 1088 uint inct_cnt = 0; 1089 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1090 1091 // Count and start of calls 1092 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1093 1094 uint return_offset = 0; 1095 int nop_size = (new MachNopNode())->size(_regalloc); 1096 1097 int previous_offset = 0; 1098 int current_offset = 0; 1099 int last_call_offset = -1; 1100 int last_avoid_back_to_back_offset = -1; 1101 #ifdef ASSERT 1102 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1103 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1104 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1105 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1106 #endif 1107 1108 // Create an array of unused labels, one for each basic block, if printing is enabled 1109 #ifndef PRODUCT 1110 int *node_offsets = NULL; 1111 uint node_offset_limit = unique(); 1112 1113 if (print_assembly()) 1114 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1115 #endif 1116 1117 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1118 1119 // Emit the constant table. 1120 if (has_mach_constant_base_node()) { 1121 constant_table().emit(*cb); 1122 } 1123 1124 // Create an array of labels, one for each basic block 1125 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1126 for (uint i=0; i <= nblocks; i++) { 1127 blk_labels[i].init(); 1128 } 1129 1130 // ------------------ 1131 // Now fill in the code buffer 1132 Node *delay_slot = NULL; 1133 1134 for (uint i = 0; i < nblocks; i++) { 1135 Block* block = _cfg->get_block(i); 1136 Node* head = block->head(); 1137 1138 // If this block needs to start aligned (i.e, can be reached other 1139 // than by falling-thru from the previous block), then force the 1140 // start of a new bundle. 1141 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1142 cb->flush_bundle(true); 1143 } 1144 1145 #ifdef ASSERT 1146 if (!block->is_connector()) { 1147 stringStream st; 1148 block->dump_head(_cfg, &st); 1149 MacroAssembler(cb).block_comment(st.as_string()); 1150 } 1151 jmp_target[i] = 0; 1152 jmp_offset[i] = 0; 1153 jmp_size[i] = 0; 1154 jmp_rule[i] = 0; 1155 #endif 1156 int blk_offset = current_offset; 1157 1158 // Define the label at the beginning of the basic block 1159 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1160 1161 uint last_inst = block->number_of_nodes(); 1162 1163 // Emit block normally, except for last instruction. 1164 // Emit means "dump code bits into code buffer". 1165 for (uint j = 0; j<last_inst; j++) { 1166 1167 // Get the node 1168 Node* n = block->get_node(j); 1169 1170 // See if delay slots are supported 1171 if (valid_bundle_info(n) && 1172 node_bundling(n)->used_in_unconditional_delay()) { 1173 assert(delay_slot == NULL, "no use of delay slot node"); 1174 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1175 1176 delay_slot = n; 1177 continue; 1178 } 1179 1180 // If this starts a new instruction group, then flush the current one 1181 // (but allow split bundles) 1182 if (Pipeline::requires_bundling() && starts_bundle(n)) 1183 cb->flush_bundle(false); 1184 1185 // Special handling for SafePoint/Call Nodes 1186 bool is_mcall = false; 1187 if (n->is_Mach()) { 1188 MachNode *mach = n->as_Mach(); 1189 is_mcall = n->is_MachCall(); 1190 bool is_sfn = n->is_MachSafePoint(); 1191 1192 // If this requires all previous instructions be flushed, then do so 1193 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1194 cb->flush_bundle(true); 1195 current_offset = cb->insts_size(); 1196 } 1197 1198 // A padding may be needed again since a previous instruction 1199 // could be moved to delay slot. 1200 1201 // align the instruction if necessary 1202 int padding = mach->compute_padding(current_offset); 1203 // Make sure safepoint node for polling is distinct from a call's 1204 // return by adding a nop if needed. 1205 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1206 padding = nop_size; 1207 } 1208 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1209 current_offset == last_avoid_back_to_back_offset) { 1210 // Avoid back to back some instructions. 1211 padding = nop_size; 1212 } 1213 1214 if (padding > 0) { 1215 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1216 int nops_cnt = padding / nop_size; 1217 MachNode *nop = new MachNopNode(nops_cnt); 1218 block->insert_node(nop, j++); 1219 last_inst++; 1220 _cfg->map_node_to_block(nop, block); 1221 // Ensure enough space. 1222 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1223 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1224 C->record_failure("CodeCache is full"); 1225 return; 1226 } 1227 nop->emit(*cb, _regalloc); 1228 cb->flush_bundle(true); 1229 current_offset = cb->insts_size(); 1230 } 1231 1232 // Remember the start of the last call in a basic block 1233 if (is_mcall) { 1234 MachCallNode *mcall = mach->as_MachCall(); 1235 1236 // This destination address is NOT PC-relative 1237 mcall->method_set((intptr_t)mcall->entry_point()); 1238 1239 // Save the return address 1240 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1241 1242 if (mcall->is_MachCallLeaf()) { 1243 is_mcall = false; 1244 is_sfn = false; 1245 } 1246 } 1247 1248 // sfn will be valid whenever mcall is valid now because of inheritance 1249 if (is_sfn || is_mcall) { 1250 1251 // Handle special safepoint nodes for synchronization 1252 if (!is_mcall) { 1253 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1254 // !!!!! Stubs only need an oopmap right now, so bail out 1255 if (sfn->jvms()->method() == NULL) { 1256 // Write the oopmap directly to the code blob??!! 1257 continue; 1258 } 1259 } // End synchronization 1260 1261 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1262 current_offset); 1263 Process_OopMap_Node(mach, current_offset); 1264 } // End if safepoint 1265 1266 // If this is a null check, then add the start of the previous instruction to the list 1267 else if( mach->is_MachNullCheck() ) { 1268 inct_starts[inct_cnt++] = previous_offset; 1269 } 1270 1271 // If this is a branch, then fill in the label with the target BB's label 1272 else if (mach->is_MachBranch()) { 1273 // This requires the TRUE branch target be in succs[0] 1274 uint block_num = block->non_connector_successor(0)->_pre_order; 1275 1276 // Try to replace long branch if delay slot is not used, 1277 // it is mostly for back branches since forward branch's 1278 // distance is not updated yet. 1279 bool delay_slot_is_used = valid_bundle_info(n) && 1280 node_bundling(n)->use_unconditional_delay(); 1281 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1282 assert(delay_slot == NULL, "not expecting delay slot node"); 1283 int br_size = n->size(_regalloc); 1284 int offset = blk_starts[block_num] - current_offset; 1285 if (block_num >= i) { 1286 // Current and following block's offset are not 1287 // finalized yet, adjust distance by the difference 1288 // between calculated and final offsets of current block. 1289 offset -= (blk_starts[i] - blk_offset); 1290 } 1291 // In the following code a nop could be inserted before 1292 // the branch which will increase the backward distance. 1293 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1294 if (needs_padding && offset <= 0) 1295 offset -= nop_size; 1296 1297 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1298 // We've got a winner. Replace this branch. 1299 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1300 1301 // Update the jmp_size. 1302 int new_size = replacement->size(_regalloc); 1303 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1304 // Insert padding between avoid_back_to_back branches. 1305 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1306 MachNode *nop = new MachNopNode(); 1307 block->insert_node(nop, j++); 1308 _cfg->map_node_to_block(nop, block); 1309 last_inst++; 1310 nop->emit(*cb, _regalloc); 1311 cb->flush_bundle(true); 1312 current_offset = cb->insts_size(); 1313 } 1314 #ifdef ASSERT 1315 jmp_target[i] = block_num; 1316 jmp_offset[i] = current_offset - blk_offset; 1317 jmp_size[i] = new_size; 1318 jmp_rule[i] = mach->rule(); 1319 #endif 1320 block->map_node(replacement, j); 1321 mach->subsume_by(replacement, C); 1322 n = replacement; 1323 mach = replacement; 1324 } 1325 } 1326 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1327 } else if (mach->ideal_Opcode() == Op_Jump) { 1328 for (uint h = 0; h < block->_num_succs; h++) { 1329 Block* succs_block = block->_succs[h]; 1330 for (uint j = 1; j < succs_block->num_preds(); j++) { 1331 Node* jpn = succs_block->pred(j); 1332 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1333 uint block_num = succs_block->non_connector()->_pre_order; 1334 Label *blkLabel = &blk_labels[block_num]; 1335 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1336 } 1337 } 1338 } 1339 } 1340 #ifdef ASSERT 1341 // Check that oop-store precedes the card-mark 1342 else if (mach->ideal_Opcode() == Op_StoreCM) { 1343 uint storeCM_idx = j; 1344 int count = 0; 1345 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1346 Node *oop_store = mach->in(prec); // Precedence edge 1347 if (oop_store == NULL) continue; 1348 count++; 1349 uint i4; 1350 for (i4 = 0; i4 < last_inst; ++i4) { 1351 if (block->get_node(i4) == oop_store) { 1352 break; 1353 } 1354 } 1355 // Note: This test can provide a false failure if other precedence 1356 // edges have been added to the storeCMNode. 1357 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1358 } 1359 assert(count > 0, "storeCM expects at least one precedence edge"); 1360 } 1361 #endif 1362 else if (!n->is_Proj()) { 1363 // Remember the beginning of the previous instruction, in case 1364 // it's followed by a flag-kill and a null-check. Happens on 1365 // Intel all the time, with add-to-memory kind of opcodes. 1366 previous_offset = current_offset; 1367 } 1368 1369 // Not an else-if! 1370 // If this is a trap based cmp then add its offset to the list. 1371 if (mach->is_TrapBasedCheckNode()) { 1372 inct_starts[inct_cnt++] = current_offset; 1373 } 1374 } 1375 1376 // Verify that there is sufficient space remaining 1377 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1378 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1379 C->record_failure("CodeCache is full"); 1380 return; 1381 } 1382 1383 // Save the offset for the listing 1384 #ifndef PRODUCT 1385 if (node_offsets && n->_idx < node_offset_limit) 1386 node_offsets[n->_idx] = cb->insts_size(); 1387 #endif 1388 1389 // "Normal" instruction case 1390 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1391 n->emit(*cb, _regalloc); 1392 current_offset = cb->insts_size(); 1393 1394 // Above we only verified that there is enough space in the instruction section. 1395 // However, the instruction may emit stubs that cause code buffer expansion. 1396 // Bail out here if expansion failed due to a lack of code cache space. 1397 if (failing()) { 1398 return; 1399 } 1400 1401 #ifdef ASSERT 1402 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1403 n->dump(); 1404 assert(false, "wrong size of mach node"); 1405 } 1406 #endif 1407 non_safepoints.observe_instruction(n, current_offset); 1408 1409 // mcall is last "call" that can be a safepoint 1410 // record it so we can see if a poll will directly follow it 1411 // in which case we'll need a pad to make the PcDesc sites unique 1412 // see 5010568. This can be slightly inaccurate but conservative 1413 // in the case that return address is not actually at current_offset. 1414 // This is a small price to pay. 1415 1416 if (is_mcall) { 1417 last_call_offset = current_offset; 1418 } 1419 1420 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1421 // Avoid back to back some instructions. 1422 last_avoid_back_to_back_offset = current_offset; 1423 } 1424 1425 // See if this instruction has a delay slot 1426 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1427 assert(delay_slot != NULL, "expecting delay slot node"); 1428 1429 // Back up 1 instruction 1430 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1431 1432 // Save the offset for the listing 1433 #ifndef PRODUCT 1434 if (node_offsets && delay_slot->_idx < node_offset_limit) 1435 node_offsets[delay_slot->_idx] = cb->insts_size(); 1436 #endif 1437 1438 // Support a SafePoint in the delay slot 1439 if (delay_slot->is_MachSafePoint()) { 1440 MachNode *mach = delay_slot->as_Mach(); 1441 // !!!!! Stubs only need an oopmap right now, so bail out 1442 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1443 // Write the oopmap directly to the code blob??!! 1444 delay_slot = NULL; 1445 continue; 1446 } 1447 1448 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1449 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1450 adjusted_offset); 1451 // Generate an OopMap entry 1452 Process_OopMap_Node(mach, adjusted_offset); 1453 } 1454 1455 // Insert the delay slot instruction 1456 delay_slot->emit(*cb, _regalloc); 1457 1458 // Don't reuse it 1459 delay_slot = NULL; 1460 } 1461 1462 } // End for all instructions in block 1463 1464 // If the next block is the top of a loop, pad this block out to align 1465 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1466 if (i < nblocks-1) { 1467 Block *nb = _cfg->get_block(i + 1); 1468 int padding = nb->alignment_padding(current_offset); 1469 if( padding > 0 ) { 1470 MachNode *nop = new MachNopNode(padding / nop_size); 1471 block->insert_node(nop, block->number_of_nodes()); 1472 _cfg->map_node_to_block(nop, block); 1473 nop->emit(*cb, _regalloc); 1474 current_offset = cb->insts_size(); 1475 } 1476 } 1477 // Verify that the distance for generated before forward 1478 // short branches is still valid. 1479 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1480 1481 // Save new block start offset 1482 blk_starts[i] = blk_offset; 1483 } // End of for all blocks 1484 blk_starts[nblocks] = current_offset; 1485 1486 non_safepoints.flush_at_end(); 1487 1488 // Offset too large? 1489 if (failing()) return; 1490 1491 // Define a pseudo-label at the end of the code 1492 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1493 1494 // Compute the size of the first block 1495 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1496 1497 #ifdef ASSERT 1498 for (uint i = 0; i < nblocks; i++) { // For all blocks 1499 if (jmp_target[i] != 0) { 1500 int br_size = jmp_size[i]; 1501 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1502 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1503 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 1504 assert(false, "Displacement too large for short jmp"); 1505 } 1506 } 1507 } 1508 #endif 1509 1510 #ifndef PRODUCT 1511 // Information on the size of the method, without the extraneous code 1512 Scheduling::increment_method_size(cb->insts_size()); 1513 #endif 1514 1515 // ------------------ 1516 // Fill in exception table entries. 1517 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1518 1519 // Only java methods have exception handlers and deopt handlers 1520 // class HandlerImpl is platform-specific and defined in the *.ad files. 1521 if (_method) { 1522 // Emit the exception handler code. 1523 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1524 if (failing()) { 1525 return; // CodeBuffer::expand failed 1526 } 1527 // Emit the deopt handler code. 1528 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1529 1530 // Emit the MethodHandle deopt handler code (if required). 1531 if (has_method_handle_invokes() && !failing()) { 1532 // We can use the same code as for the normal deopt handler, we 1533 // just need a different entry point address. 1534 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1535 } 1536 } 1537 1538 // One last check for failed CodeBuffer::expand: 1539 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1540 C->record_failure("CodeCache is full"); 1541 return; 1542 } 1543 1544 #ifndef PRODUCT 1545 // Dump the assembly code, including basic-block numbers 1546 if (print_assembly()) { 1547 ttyLocker ttyl; // keep the following output all in one block 1548 if (!VMThread::should_terminate()) { // test this under the tty lock 1549 // This output goes directly to the tty, not the compiler log. 1550 // To enable tools to match it up with the compilation activity, 1551 // be sure to tag this tty output with the compile ID. 1552 if (xtty != NULL) { 1553 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1554 is_osr_compilation() ? " compile_kind='osr'" : 1555 ""); 1556 } 1557 if (method() != NULL) { 1558 method()->print_metadata(); 1559 } 1560 dump_asm(node_offsets, node_offset_limit); 1561 if (xtty != NULL) { 1562 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock. 1563 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done 1564 // thread safe 1565 ttyLocker ttyl2; 1566 xtty->tail("opto_assembly"); 1567 } 1568 } 1569 } 1570 #endif 1571 1572 } 1573 1574 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1575 _inc_table.set_size(cnt); 1576 1577 uint inct_cnt = 0; 1578 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1579 Block* block = _cfg->get_block(i); 1580 Node *n = NULL; 1581 int j; 1582 1583 // Find the branch; ignore trailing NOPs. 1584 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1585 n = block->get_node(j); 1586 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1587 break; 1588 } 1589 } 1590 1591 // If we didn't find anything, continue 1592 if (j < 0) { 1593 continue; 1594 } 1595 1596 // Compute ExceptionHandlerTable subtable entry and add it 1597 // (skip empty blocks) 1598 if (n->is_Catch()) { 1599 1600 // Get the offset of the return from the call 1601 uint call_return = call_returns[block->_pre_order]; 1602 #ifdef ASSERT 1603 assert( call_return > 0, "no call seen for this basic block" ); 1604 while (block->get_node(--j)->is_MachProj()) ; 1605 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1606 #endif 1607 // last instruction is a CatchNode, find it's CatchProjNodes 1608 int nof_succs = block->_num_succs; 1609 // allocate space 1610 GrowableArray<intptr_t> handler_bcis(nof_succs); 1611 GrowableArray<intptr_t> handler_pcos(nof_succs); 1612 // iterate through all successors 1613 for (int j = 0; j < nof_succs; j++) { 1614 Block* s = block->_succs[j]; 1615 bool found_p = false; 1616 for (uint k = 1; k < s->num_preds(); k++) { 1617 Node* pk = s->pred(k); 1618 if (pk->is_CatchProj() && pk->in(0) == n) { 1619 const CatchProjNode* p = pk->as_CatchProj(); 1620 found_p = true; 1621 // add the corresponding handler bci & pco information 1622 if (p->_con != CatchProjNode::fall_through_index) { 1623 // p leads to an exception handler (and is not fall through) 1624 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1625 // no duplicates, please 1626 if (!handler_bcis.contains(p->handler_bci())) { 1627 uint block_num = s->non_connector()->_pre_order; 1628 handler_bcis.append(p->handler_bci()); 1629 handler_pcos.append(blk_labels[block_num].loc_pos()); 1630 } 1631 } 1632 } 1633 } 1634 assert(found_p, "no matching predecessor found"); 1635 // Note: Due to empty block removal, one block may have 1636 // several CatchProj inputs, from the same Catch. 1637 } 1638 1639 // Set the offset of the return from the call 1640 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1641 continue; 1642 } 1643 1644 // Handle implicit null exception table updates 1645 if (n->is_MachNullCheck()) { 1646 uint block_num = block->non_connector_successor(0)->_pre_order; 1647 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1648 continue; 1649 } 1650 // Handle implicit exception table updates: trap instructions. 1651 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1652 uint block_num = block->non_connector_successor(0)->_pre_order; 1653 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1654 continue; 1655 } 1656 } // End of for all blocks fill in exception table entries 1657 } 1658 1659 // Static Variables 1660 #ifndef PRODUCT 1661 uint Scheduling::_total_nop_size = 0; 1662 uint Scheduling::_total_method_size = 0; 1663 uint Scheduling::_total_branches = 0; 1664 uint Scheduling::_total_unconditional_delays = 0; 1665 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1666 #endif 1667 1668 // Initializer for class Scheduling 1669 1670 Scheduling::Scheduling(Arena *arena, Compile &compile) 1671 : _arena(arena), 1672 _cfg(compile.cfg()), 1673 _regalloc(compile.regalloc()), 1674 _reg_node(arena), 1675 _bundle_instr_count(0), 1676 _bundle_cycle_number(0), 1677 _scheduled(arena), 1678 _available(arena), 1679 _next_node(NULL), 1680 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1681 _pinch_free_list(arena) 1682 #ifndef PRODUCT 1683 , _branches(0) 1684 , _unconditional_delays(0) 1685 #endif 1686 { 1687 // Create a MachNopNode 1688 _nop = new MachNopNode(); 1689 1690 // Now that the nops are in the array, save the count 1691 // (but allow entries for the nops) 1692 _node_bundling_limit = compile.unique(); 1693 uint node_max = _regalloc->node_regs_max_index(); 1694 1695 compile.set_node_bundling_limit(_node_bundling_limit); 1696 1697 // This one is persistent within the Compile class 1698 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1699 1700 // Allocate space for fixed-size arrays 1701 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1702 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1703 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1704 1705 // Clear the arrays 1706 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1707 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1708 memset(_uses, 0, node_max * sizeof(short)); 1709 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1710 1711 // Clear the bundling information 1712 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1713 1714 // Get the last node 1715 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1716 1717 _next_node = block->get_node(block->number_of_nodes() - 1); 1718 } 1719 1720 #ifndef PRODUCT 1721 // Scheduling destructor 1722 Scheduling::~Scheduling() { 1723 _total_branches += _branches; 1724 _total_unconditional_delays += _unconditional_delays; 1725 } 1726 #endif 1727 1728 // Step ahead "i" cycles 1729 void Scheduling::step(uint i) { 1730 1731 Bundle *bundle = node_bundling(_next_node); 1732 bundle->set_starts_bundle(); 1733 1734 // Update the bundle record, but leave the flags information alone 1735 if (_bundle_instr_count > 0) { 1736 bundle->set_instr_count(_bundle_instr_count); 1737 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1738 } 1739 1740 // Update the state information 1741 _bundle_instr_count = 0; 1742 _bundle_cycle_number += i; 1743 _bundle_use.step(i); 1744 } 1745 1746 void Scheduling::step_and_clear() { 1747 Bundle *bundle = node_bundling(_next_node); 1748 bundle->set_starts_bundle(); 1749 1750 // Update the bundle record 1751 if (_bundle_instr_count > 0) { 1752 bundle->set_instr_count(_bundle_instr_count); 1753 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1754 1755 _bundle_cycle_number += 1; 1756 } 1757 1758 // Clear the bundling information 1759 _bundle_instr_count = 0; 1760 _bundle_use.reset(); 1761 1762 memcpy(_bundle_use_elements, 1763 Pipeline_Use::elaborated_elements, 1764 sizeof(Pipeline_Use::elaborated_elements)); 1765 } 1766 1767 // Perform instruction scheduling and bundling over the sequence of 1768 // instructions in backwards order. 1769 void Compile::ScheduleAndBundle() { 1770 1771 // Don't optimize this if it isn't a method 1772 if (!_method) 1773 return; 1774 1775 // Don't optimize this if scheduling is disabled 1776 if (!do_scheduling()) 1777 return; 1778 1779 // Scheduling code works only with pairs (16 bytes) maximum. 1780 if (max_vector_size() > 16) 1781 return; 1782 1783 TracePhase tp("isched", &timers[_t_instrSched]); 1784 1785 // Create a data structure for all the scheduling information 1786 Scheduling scheduling(Thread::current()->resource_area(), *this); 1787 1788 // Walk backwards over each basic block, computing the needed alignment 1789 // Walk over all the basic blocks 1790 scheduling.DoScheduling(); 1791 } 1792 1793 // Compute the latency of all the instructions. This is fairly simple, 1794 // because we already have a legal ordering. Walk over the instructions 1795 // from first to last, and compute the latency of the instruction based 1796 // on the latency of the preceding instruction(s). 1797 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1798 #ifndef PRODUCT 1799 if (_cfg->C->trace_opto_output()) 1800 tty->print("# -> ComputeLocalLatenciesForward\n"); 1801 #endif 1802 1803 // Walk over all the schedulable instructions 1804 for( uint j=_bb_start; j < _bb_end; j++ ) { 1805 1806 // This is a kludge, forcing all latency calculations to start at 1. 1807 // Used to allow latency 0 to force an instruction to the beginning 1808 // of the bb 1809 uint latency = 1; 1810 Node *use = bb->get_node(j); 1811 uint nlen = use->len(); 1812 1813 // Walk over all the inputs 1814 for ( uint k=0; k < nlen; k++ ) { 1815 Node *def = use->in(k); 1816 if (!def) 1817 continue; 1818 1819 uint l = _node_latency[def->_idx] + use->latency(k); 1820 if (latency < l) 1821 latency = l; 1822 } 1823 1824 _node_latency[use->_idx] = latency; 1825 1826 #ifndef PRODUCT 1827 if (_cfg->C->trace_opto_output()) { 1828 tty->print("# latency %4d: ", latency); 1829 use->dump(); 1830 } 1831 #endif 1832 } 1833 1834 #ifndef PRODUCT 1835 if (_cfg->C->trace_opto_output()) 1836 tty->print("# <- ComputeLocalLatenciesForward\n"); 1837 #endif 1838 1839 } // end ComputeLocalLatenciesForward 1840 1841 // See if this node fits into the present instruction bundle 1842 bool Scheduling::NodeFitsInBundle(Node *n) { 1843 uint n_idx = n->_idx; 1844 1845 // If this is the unconditional delay instruction, then it fits 1846 if (n == _unconditional_delay_slot) { 1847 #ifndef PRODUCT 1848 if (_cfg->C->trace_opto_output()) 1849 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1850 #endif 1851 return (true); 1852 } 1853 1854 // If the node cannot be scheduled this cycle, skip it 1855 if (_current_latency[n_idx] > _bundle_cycle_number) { 1856 #ifndef PRODUCT 1857 if (_cfg->C->trace_opto_output()) 1858 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1859 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1860 #endif 1861 return (false); 1862 } 1863 1864 const Pipeline *node_pipeline = n->pipeline(); 1865 1866 uint instruction_count = node_pipeline->instructionCount(); 1867 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1868 instruction_count = 0; 1869 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1870 instruction_count++; 1871 1872 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1873 #ifndef PRODUCT 1874 if (_cfg->C->trace_opto_output()) 1875 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1876 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1877 #endif 1878 return (false); 1879 } 1880 1881 // Don't allow non-machine nodes to be handled this way 1882 if (!n->is_Mach() && instruction_count == 0) 1883 return (false); 1884 1885 // See if there is any overlap 1886 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1887 1888 if (delay > 0) { 1889 #ifndef PRODUCT 1890 if (_cfg->C->trace_opto_output()) 1891 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1892 #endif 1893 return false; 1894 } 1895 1896 #ifndef PRODUCT 1897 if (_cfg->C->trace_opto_output()) 1898 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1899 #endif 1900 1901 return true; 1902 } 1903 1904 Node * Scheduling::ChooseNodeToBundle() { 1905 uint siz = _available.size(); 1906 1907 if (siz == 0) { 1908 1909 #ifndef PRODUCT 1910 if (_cfg->C->trace_opto_output()) 1911 tty->print("# ChooseNodeToBundle: NULL\n"); 1912 #endif 1913 return (NULL); 1914 } 1915 1916 // Fast path, if only 1 instruction in the bundle 1917 if (siz == 1) { 1918 #ifndef PRODUCT 1919 if (_cfg->C->trace_opto_output()) { 1920 tty->print("# ChooseNodeToBundle (only 1): "); 1921 _available[0]->dump(); 1922 } 1923 #endif 1924 return (_available[0]); 1925 } 1926 1927 // Don't bother, if the bundle is already full 1928 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1929 for ( uint i = 0; i < siz; i++ ) { 1930 Node *n = _available[i]; 1931 1932 // Skip projections, we'll handle them another way 1933 if (n->is_Proj()) 1934 continue; 1935 1936 // This presupposed that instructions are inserted into the 1937 // available list in a legality order; i.e. instructions that 1938 // must be inserted first are at the head of the list 1939 if (NodeFitsInBundle(n)) { 1940 #ifndef PRODUCT 1941 if (_cfg->C->trace_opto_output()) { 1942 tty->print("# ChooseNodeToBundle: "); 1943 n->dump(); 1944 } 1945 #endif 1946 return (n); 1947 } 1948 } 1949 } 1950 1951 // Nothing fits in this bundle, choose the highest priority 1952 #ifndef PRODUCT 1953 if (_cfg->C->trace_opto_output()) { 1954 tty->print("# ChooseNodeToBundle: "); 1955 _available[0]->dump(); 1956 } 1957 #endif 1958 1959 return _available[0]; 1960 } 1961 1962 void Scheduling::AddNodeToAvailableList(Node *n) { 1963 assert( !n->is_Proj(), "projections never directly made available" ); 1964 #ifndef PRODUCT 1965 if (_cfg->C->trace_opto_output()) { 1966 tty->print("# AddNodeToAvailableList: "); 1967 n->dump(); 1968 } 1969 #endif 1970 1971 int latency = _current_latency[n->_idx]; 1972 1973 // Insert in latency order (insertion sort) 1974 uint i; 1975 for ( i=0; i < _available.size(); i++ ) 1976 if (_current_latency[_available[i]->_idx] > latency) 1977 break; 1978 1979 // Special Check for compares following branches 1980 if( n->is_Mach() && _scheduled.size() > 0 ) { 1981 int op = n->as_Mach()->ideal_Opcode(); 1982 Node *last = _scheduled[0]; 1983 if( last->is_MachIf() && last->in(1) == n && 1984 ( op == Op_CmpI || 1985 op == Op_CmpU || 1986 op == Op_CmpP || 1987 op == Op_CmpF || 1988 op == Op_CmpD || 1989 op == Op_CmpL ) ) { 1990 1991 // Recalculate position, moving to front of same latency 1992 for ( i=0 ; i < _available.size(); i++ ) 1993 if (_current_latency[_available[i]->_idx] >= latency) 1994 break; 1995 } 1996 } 1997 1998 // Insert the node in the available list 1999 _available.insert(i, n); 2000 2001 #ifndef PRODUCT 2002 if (_cfg->C->trace_opto_output()) 2003 dump_available(); 2004 #endif 2005 } 2006 2007 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2008 for ( uint i=0; i < n->len(); i++ ) { 2009 Node *def = n->in(i); 2010 if (!def) continue; 2011 if( def->is_Proj() ) // If this is a machine projection, then 2012 def = def->in(0); // propagate usage thru to the base instruction 2013 2014 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2015 continue; 2016 } 2017 2018 // Compute the latency 2019 uint l = _bundle_cycle_number + n->latency(i); 2020 if (_current_latency[def->_idx] < l) 2021 _current_latency[def->_idx] = l; 2022 2023 // If this does not have uses then schedule it 2024 if ((--_uses[def->_idx]) == 0) 2025 AddNodeToAvailableList(def); 2026 } 2027 } 2028 2029 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2030 #ifndef PRODUCT 2031 if (_cfg->C->trace_opto_output()) { 2032 tty->print("# AddNodeToBundle: "); 2033 n->dump(); 2034 } 2035 #endif 2036 2037 // Remove this from the available list 2038 uint i; 2039 for (i = 0; i < _available.size(); i++) 2040 if (_available[i] == n) 2041 break; 2042 assert(i < _available.size(), "entry in _available list not found"); 2043 _available.remove(i); 2044 2045 // See if this fits in the current bundle 2046 const Pipeline *node_pipeline = n->pipeline(); 2047 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2048 2049 // Check for instructions to be placed in the delay slot. We 2050 // do this before we actually schedule the current instruction, 2051 // because the delay slot follows the current instruction. 2052 if (Pipeline::_branch_has_delay_slot && 2053 node_pipeline->hasBranchDelay() && 2054 !_unconditional_delay_slot) { 2055 2056 uint siz = _available.size(); 2057 2058 // Conditional branches can support an instruction that 2059 // is unconditionally executed and not dependent by the 2060 // branch, OR a conditionally executed instruction if 2061 // the branch is taken. In practice, this means that 2062 // the first instruction at the branch target is 2063 // copied to the delay slot, and the branch goes to 2064 // the instruction after that at the branch target 2065 if ( n->is_MachBranch() ) { 2066 2067 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2068 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2069 2070 #ifndef PRODUCT 2071 _branches++; 2072 #endif 2073 2074 // At least 1 instruction is on the available list 2075 // that is not dependent on the branch 2076 for (uint i = 0; i < siz; i++) { 2077 Node *d = _available[i]; 2078 const Pipeline *avail_pipeline = d->pipeline(); 2079 2080 // Don't allow safepoints in the branch shadow, that will 2081 // cause a number of difficulties 2082 if ( avail_pipeline->instructionCount() == 1 && 2083 !avail_pipeline->hasMultipleBundles() && 2084 !avail_pipeline->hasBranchDelay() && 2085 Pipeline::instr_has_unit_size() && 2086 d->size(_regalloc) == Pipeline::instr_unit_size() && 2087 NodeFitsInBundle(d) && 2088 !node_bundling(d)->used_in_delay()) { 2089 2090 if (d->is_Mach() && !d->is_MachSafePoint()) { 2091 // A node that fits in the delay slot was found, so we need to 2092 // set the appropriate bits in the bundle pipeline information so 2093 // that it correctly indicates resource usage. Later, when we 2094 // attempt to add this instruction to the bundle, we will skip 2095 // setting the resource usage. 2096 _unconditional_delay_slot = d; 2097 node_bundling(n)->set_use_unconditional_delay(); 2098 node_bundling(d)->set_used_in_unconditional_delay(); 2099 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2100 _current_latency[d->_idx] = _bundle_cycle_number; 2101 _next_node = d; 2102 ++_bundle_instr_count; 2103 #ifndef PRODUCT 2104 _unconditional_delays++; 2105 #endif 2106 break; 2107 } 2108 } 2109 } 2110 } 2111 2112 // No delay slot, add a nop to the usage 2113 if (!_unconditional_delay_slot) { 2114 // See if adding an instruction in the delay slot will overflow 2115 // the bundle. 2116 if (!NodeFitsInBundle(_nop)) { 2117 #ifndef PRODUCT 2118 if (_cfg->C->trace_opto_output()) 2119 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2120 #endif 2121 step(1); 2122 } 2123 2124 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2125 _next_node = _nop; 2126 ++_bundle_instr_count; 2127 } 2128 2129 // See if the instruction in the delay slot requires a 2130 // step of the bundles 2131 if (!NodeFitsInBundle(n)) { 2132 #ifndef PRODUCT 2133 if (_cfg->C->trace_opto_output()) 2134 tty->print("# *** STEP(branch won't fit) ***\n"); 2135 #endif 2136 // Update the state information 2137 _bundle_instr_count = 0; 2138 _bundle_cycle_number += 1; 2139 _bundle_use.step(1); 2140 } 2141 } 2142 2143 // Get the number of instructions 2144 uint instruction_count = node_pipeline->instructionCount(); 2145 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2146 instruction_count = 0; 2147 2148 // Compute the latency information 2149 uint delay = 0; 2150 2151 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2152 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2153 if (relative_latency < 0) 2154 relative_latency = 0; 2155 2156 delay = _bundle_use.full_latency(relative_latency, node_usage); 2157 2158 // Does not fit in this bundle, start a new one 2159 if (delay > 0) { 2160 step(delay); 2161 2162 #ifndef PRODUCT 2163 if (_cfg->C->trace_opto_output()) 2164 tty->print("# *** STEP(%d) ***\n", delay); 2165 #endif 2166 } 2167 } 2168 2169 // If this was placed in the delay slot, ignore it 2170 if (n != _unconditional_delay_slot) { 2171 2172 if (delay == 0) { 2173 if (node_pipeline->hasMultipleBundles()) { 2174 #ifndef PRODUCT 2175 if (_cfg->C->trace_opto_output()) 2176 tty->print("# *** STEP(multiple instructions) ***\n"); 2177 #endif 2178 step(1); 2179 } 2180 2181 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2182 #ifndef PRODUCT 2183 if (_cfg->C->trace_opto_output()) 2184 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2185 instruction_count + _bundle_instr_count, 2186 Pipeline::_max_instrs_per_cycle); 2187 #endif 2188 step(1); 2189 } 2190 } 2191 2192 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2193 _bundle_instr_count++; 2194 2195 // Set the node's latency 2196 _current_latency[n->_idx] = _bundle_cycle_number; 2197 2198 // Now merge the functional unit information 2199 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2200 _bundle_use.add_usage(node_usage); 2201 2202 // Increment the number of instructions in this bundle 2203 _bundle_instr_count += instruction_count; 2204 2205 // Remember this node for later 2206 if (n->is_Mach()) 2207 _next_node = n; 2208 } 2209 2210 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2211 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2212 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2213 // into the block. All other scheduled nodes get put in the schedule here. 2214 int op = n->Opcode(); 2215 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2216 (op != Op_Node && // Not an unused antidepedence node and 2217 // not an unallocated boxlock 2218 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2219 2220 // Push any trailing projections 2221 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2222 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2223 Node *foi = n->fast_out(i); 2224 if( foi->is_Proj() ) 2225 _scheduled.push(foi); 2226 } 2227 } 2228 2229 // Put the instruction in the schedule list 2230 _scheduled.push(n); 2231 } 2232 2233 #ifndef PRODUCT 2234 if (_cfg->C->trace_opto_output()) 2235 dump_available(); 2236 #endif 2237 2238 // Walk all the definitions, decrementing use counts, and 2239 // if a definition has a 0 use count, place it in the available list. 2240 DecrementUseCounts(n,bb); 2241 } 2242 2243 // This method sets the use count within a basic block. We will ignore all 2244 // uses outside the current basic block. As we are doing a backwards walk, 2245 // any node we reach that has a use count of 0 may be scheduled. This also 2246 // avoids the problem of cyclic references from phi nodes, as long as phi 2247 // nodes are at the front of the basic block. This method also initializes 2248 // the available list to the set of instructions that have no uses within this 2249 // basic block. 2250 void Scheduling::ComputeUseCount(const Block *bb) { 2251 #ifndef PRODUCT 2252 if (_cfg->C->trace_opto_output()) 2253 tty->print("# -> ComputeUseCount\n"); 2254 #endif 2255 2256 // Clear the list of available and scheduled instructions, just in case 2257 _available.clear(); 2258 _scheduled.clear(); 2259 2260 // No delay slot specified 2261 _unconditional_delay_slot = NULL; 2262 2263 #ifdef ASSERT 2264 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2265 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2266 #endif 2267 2268 // Force the _uses count to never go to zero for unscheduable pieces 2269 // of the block 2270 for( uint k = 0; k < _bb_start; k++ ) 2271 _uses[bb->get_node(k)->_idx] = 1; 2272 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2273 _uses[bb->get_node(l)->_idx] = 1; 2274 2275 // Iterate backwards over the instructions in the block. Don't count the 2276 // branch projections at end or the block header instructions. 2277 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2278 Node *n = bb->get_node(j); 2279 if( n->is_Proj() ) continue; // Projections handled another way 2280 2281 // Account for all uses 2282 for ( uint k = 0; k < n->len(); k++ ) { 2283 Node *inp = n->in(k); 2284 if (!inp) continue; 2285 assert(inp != n, "no cycles allowed" ); 2286 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2287 if (inp->is_Proj()) { // Skip through Proj's 2288 inp = inp->in(0); 2289 } 2290 ++_uses[inp->_idx]; // Count 1 block-local use 2291 } 2292 } 2293 2294 // If this instruction has a 0 use count, then it is available 2295 if (!_uses[n->_idx]) { 2296 _current_latency[n->_idx] = _bundle_cycle_number; 2297 AddNodeToAvailableList(n); 2298 } 2299 2300 #ifndef PRODUCT 2301 if (_cfg->C->trace_opto_output()) { 2302 tty->print("# uses: %3d: ", _uses[n->_idx]); 2303 n->dump(); 2304 } 2305 #endif 2306 } 2307 2308 #ifndef PRODUCT 2309 if (_cfg->C->trace_opto_output()) 2310 tty->print("# <- ComputeUseCount\n"); 2311 #endif 2312 } 2313 2314 // This routine performs scheduling on each basic block in reverse order, 2315 // using instruction latencies and taking into account function unit 2316 // availability. 2317 void Scheduling::DoScheduling() { 2318 #ifndef PRODUCT 2319 if (_cfg->C->trace_opto_output()) 2320 tty->print("# -> DoScheduling\n"); 2321 #endif 2322 2323 Block *succ_bb = NULL; 2324 Block *bb; 2325 2326 // Walk over all the basic blocks in reverse order 2327 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2328 bb = _cfg->get_block(i); 2329 2330 #ifndef PRODUCT 2331 if (_cfg->C->trace_opto_output()) { 2332 tty->print("# Schedule BB#%03d (initial)\n", i); 2333 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2334 bb->get_node(j)->dump(); 2335 } 2336 } 2337 #endif 2338 2339 // On the head node, skip processing 2340 if (bb == _cfg->get_root_block()) { 2341 continue; 2342 } 2343 2344 // Skip empty, connector blocks 2345 if (bb->is_connector()) 2346 continue; 2347 2348 // If the following block is not the sole successor of 2349 // this one, then reset the pipeline information 2350 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2351 #ifndef PRODUCT 2352 if (_cfg->C->trace_opto_output()) { 2353 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2354 _next_node->_idx, _bundle_instr_count); 2355 } 2356 #endif 2357 step_and_clear(); 2358 } 2359 2360 // Leave untouched the starting instruction, any Phis, a CreateEx node 2361 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2362 _bb_end = bb->number_of_nodes()-1; 2363 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2364 Node *n = bb->get_node(_bb_start); 2365 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2366 // Also, MachIdealNodes do not get scheduled 2367 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2368 MachNode *mach = n->as_Mach(); 2369 int iop = mach->ideal_Opcode(); 2370 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2371 if( iop == Op_Con ) continue; // Do not schedule Top 2372 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2373 mach->pipeline() == MachNode::pipeline_class() && 2374 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2375 continue; 2376 break; // Funny loop structure to be sure... 2377 } 2378 // Compute last "interesting" instruction in block - last instruction we 2379 // might schedule. _bb_end points just after last schedulable inst. We 2380 // normally schedule conditional branches (despite them being forced last 2381 // in the block), because they have delay slots we can fill. Calls all 2382 // have their delay slots filled in the template expansions, so we don't 2383 // bother scheduling them. 2384 Node *last = bb->get_node(_bb_end); 2385 // Ignore trailing NOPs. 2386 while (_bb_end > 0 && last->is_Mach() && 2387 last->as_Mach()->ideal_Opcode() == Op_Con) { 2388 last = bb->get_node(--_bb_end); 2389 } 2390 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2391 if( last->is_Catch() || 2392 // Exclude unreachable path case when Halt node is in a separate block. 2393 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2394 // There must be a prior call. Skip it. 2395 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2396 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2397 } 2398 } else if( last->is_MachNullCheck() ) { 2399 // Backup so the last null-checked memory instruction is 2400 // outside the schedulable range. Skip over the nullcheck, 2401 // projection, and the memory nodes. 2402 Node *mem = last->in(1); 2403 do { 2404 _bb_end--; 2405 } while (mem != bb->get_node(_bb_end)); 2406 } else { 2407 // Set _bb_end to point after last schedulable inst. 2408 _bb_end++; 2409 } 2410 2411 assert( _bb_start <= _bb_end, "inverted block ends" ); 2412 2413 // Compute the register antidependencies for the basic block 2414 ComputeRegisterAntidependencies(bb); 2415 if (_cfg->C->failing()) return; // too many D-U pinch points 2416 2417 // Compute intra-bb latencies for the nodes 2418 ComputeLocalLatenciesForward(bb); 2419 2420 // Compute the usage within the block, and set the list of all nodes 2421 // in the block that have no uses within the block. 2422 ComputeUseCount(bb); 2423 2424 // Schedule the remaining instructions in the block 2425 while ( _available.size() > 0 ) { 2426 Node *n = ChooseNodeToBundle(); 2427 guarantee(n != NULL, "no nodes available"); 2428 AddNodeToBundle(n,bb); 2429 } 2430 2431 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2432 #ifdef ASSERT 2433 for( uint l = _bb_start; l < _bb_end; l++ ) { 2434 Node *n = bb->get_node(l); 2435 uint m; 2436 for( m = 0; m < _bb_end-_bb_start; m++ ) 2437 if( _scheduled[m] == n ) 2438 break; 2439 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2440 } 2441 #endif 2442 2443 // Now copy the instructions (in reverse order) back to the block 2444 for ( uint k = _bb_start; k < _bb_end; k++ ) 2445 bb->map_node(_scheduled[_bb_end-k-1], k); 2446 2447 #ifndef PRODUCT 2448 if (_cfg->C->trace_opto_output()) { 2449 tty->print("# Schedule BB#%03d (final)\n", i); 2450 uint current = 0; 2451 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2452 Node *n = bb->get_node(j); 2453 if( valid_bundle_info(n) ) { 2454 Bundle *bundle = node_bundling(n); 2455 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2456 tty->print("*** Bundle: "); 2457 bundle->dump(); 2458 } 2459 n->dump(); 2460 } 2461 } 2462 } 2463 #endif 2464 #ifdef ASSERT 2465 verify_good_schedule(bb,"after block local scheduling"); 2466 #endif 2467 } 2468 2469 #ifndef PRODUCT 2470 if (_cfg->C->trace_opto_output()) 2471 tty->print("# <- DoScheduling\n"); 2472 #endif 2473 2474 // Record final node-bundling array location 2475 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2476 2477 } // end DoScheduling 2478 2479 // Verify that no live-range used in the block is killed in the block by a 2480 // wrong DEF. This doesn't verify live-ranges that span blocks. 2481 2482 // Check for edge existence. Used to avoid adding redundant precedence edges. 2483 static bool edge_from_to( Node *from, Node *to ) { 2484 for( uint i=0; i<from->len(); i++ ) 2485 if( from->in(i) == to ) 2486 return true; 2487 return false; 2488 } 2489 2490 #ifdef ASSERT 2491 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2492 // Check for bad kills 2493 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2494 Node *prior_use = _reg_node[def]; 2495 if( prior_use && !edge_from_to(prior_use,n) ) { 2496 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2497 n->dump(); 2498 tty->print_cr("..."); 2499 prior_use->dump(); 2500 assert(edge_from_to(prior_use,n), "%s", msg); 2501 } 2502 _reg_node.map(def,NULL); // Kill live USEs 2503 } 2504 } 2505 2506 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2507 2508 // Zap to something reasonable for the verify code 2509 _reg_node.clear(); 2510 2511 // Walk over the block backwards. Check to make sure each DEF doesn't 2512 // kill a live value (other than the one it's supposed to). Add each 2513 // USE to the live set. 2514 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2515 Node *n = b->get_node(i); 2516 int n_op = n->Opcode(); 2517 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2518 // Fat-proj kills a slew of registers 2519 RegMask rm = n->out_RegMask();// Make local copy 2520 while( rm.is_NotEmpty() ) { 2521 OptoReg::Name kill = rm.find_first_elem(); 2522 rm.Remove(kill); 2523 verify_do_def( n, kill, msg ); 2524 } 2525 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2526 // Get DEF'd registers the normal way 2527 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2528 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2529 } 2530 2531 // Now make all USEs live 2532 for( uint i=1; i<n->req(); i++ ) { 2533 Node *def = n->in(i); 2534 assert(def != 0, "input edge required"); 2535 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2536 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2537 if( OptoReg::is_valid(reg_lo) ) { 2538 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2539 _reg_node.map(reg_lo,n); 2540 } 2541 if( OptoReg::is_valid(reg_hi) ) { 2542 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2543 _reg_node.map(reg_hi,n); 2544 } 2545 } 2546 2547 } 2548 2549 // Zap to something reasonable for the Antidependence code 2550 _reg_node.clear(); 2551 } 2552 #endif 2553 2554 // Conditionally add precedence edges. Avoid putting edges on Projs. 2555 static void add_prec_edge_from_to( Node *from, Node *to ) { 2556 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2557 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2558 from = from->in(0); 2559 } 2560 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2561 !edge_from_to( from, to ) ) // Avoid duplicate edge 2562 from->add_prec(to); 2563 } 2564 2565 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2566 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2567 return; 2568 2569 Node *pinch = _reg_node[def_reg]; // Get pinch point 2570 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2571 is_def ) { // Check for a true def (not a kill) 2572 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2573 return; 2574 } 2575 2576 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2577 debug_only( def = (Node*)0xdeadbeef; ) 2578 2579 // After some number of kills there _may_ be a later def 2580 Node *later_def = NULL; 2581 2582 // Finding a kill requires a real pinch-point. 2583 // Check for not already having a pinch-point. 2584 // Pinch points are Op_Node's. 2585 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2586 later_def = pinch; // Must be def/kill as optimistic pinch-point 2587 if ( _pinch_free_list.size() > 0) { 2588 pinch = _pinch_free_list.pop(); 2589 } else { 2590 pinch = new Node(1); // Pinch point to-be 2591 } 2592 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2593 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2594 return; 2595 } 2596 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2597 _reg_node.map(def_reg,pinch); // Record pinch-point 2598 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2599 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2600 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2601 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2602 later_def = NULL; // and no later def 2603 } 2604 pinch->set_req(0,later_def); // Hook later def so we can find it 2605 } else { // Else have valid pinch point 2606 if( pinch->in(0) ) // If there is a later-def 2607 later_def = pinch->in(0); // Get it 2608 } 2609 2610 // Add output-dependence edge from later def to kill 2611 if( later_def ) // If there is some original def 2612 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2613 2614 // See if current kill is also a use, and so is forced to be the pinch-point. 2615 if( pinch->Opcode() == Op_Node ) { 2616 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2617 for( uint i=1; i<uses->req(); i++ ) { 2618 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2619 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2620 // Yes, found a use/kill pinch-point 2621 pinch->set_req(0,NULL); // 2622 pinch->replace_by(kill); // Move anti-dep edges up 2623 pinch = kill; 2624 _reg_node.map(def_reg,pinch); 2625 return; 2626 } 2627 } 2628 } 2629 2630 // Add edge from kill to pinch-point 2631 add_prec_edge_from_to(kill,pinch); 2632 } 2633 2634 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2635 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2636 return; 2637 Node *pinch = _reg_node[use_reg]; // Get pinch point 2638 // Check for no later def_reg/kill in block 2639 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2640 // Use has to be block-local as well 2641 _cfg->get_block_for_node(use) == b) { 2642 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2643 pinch->req() == 1 ) { // pinch not yet in block? 2644 pinch->del_req(0); // yank pointer to later-def, also set flag 2645 // Insert the pinch-point in the block just after the last use 2646 b->insert_node(pinch, b->find_node(use) + 1); 2647 _bb_end++; // Increase size scheduled region in block 2648 } 2649 2650 add_prec_edge_from_to(pinch,use); 2651 } 2652 } 2653 2654 // We insert antidependences between the reads and following write of 2655 // allocated registers to prevent illegal code motion. Hopefully, the 2656 // number of added references should be fairly small, especially as we 2657 // are only adding references within the current basic block. 2658 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2659 2660 #ifdef ASSERT 2661 verify_good_schedule(b,"before block local scheduling"); 2662 #endif 2663 2664 // A valid schedule, for each register independently, is an endless cycle 2665 // of: a def, then some uses (connected to the def by true dependencies), 2666 // then some kills (defs with no uses), finally the cycle repeats with a new 2667 // def. The uses are allowed to float relative to each other, as are the 2668 // kills. No use is allowed to slide past a kill (or def). This requires 2669 // antidependencies between all uses of a single def and all kills that 2670 // follow, up to the next def. More edges are redundant, because later defs 2671 // & kills are already serialized with true or antidependencies. To keep 2672 // the edge count down, we add a 'pinch point' node if there's more than 2673 // one use or more than one kill/def. 2674 2675 // We add dependencies in one bottom-up pass. 2676 2677 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2678 2679 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2680 // register. If not, we record the DEF/KILL in _reg_node, the 2681 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2682 // "pinch point", a new Node that's in the graph but not in the block. 2683 // We put edges from the prior and current DEF/KILLs to the pinch point. 2684 // We put the pinch point in _reg_node. If there's already a pinch point 2685 // we merely add an edge from the current DEF/KILL to the pinch point. 2686 2687 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2688 // put an edge from the pinch point to the USE. 2689 2690 // To be expedient, the _reg_node array is pre-allocated for the whole 2691 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2692 // or a valid def/kill/pinch-point, or a leftover node from some prior 2693 // block. Leftover node from some prior block is treated like a NULL (no 2694 // prior def, so no anti-dependence needed). Valid def is distinguished by 2695 // it being in the current block. 2696 bool fat_proj_seen = false; 2697 uint last_safept = _bb_end-1; 2698 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2699 Node* last_safept_node = end_node; 2700 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2701 Node *n = b->get_node(i); 2702 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2703 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2704 // Fat-proj kills a slew of registers 2705 // This can add edges to 'n' and obscure whether or not it was a def, 2706 // hence the is_def flag. 2707 fat_proj_seen = true; 2708 RegMask rm = n->out_RegMask();// Make local copy 2709 while( rm.is_NotEmpty() ) { 2710 OptoReg::Name kill = rm.find_first_elem(); 2711 rm.Remove(kill); 2712 anti_do_def( b, n, kill, is_def ); 2713 } 2714 } else { 2715 // Get DEF'd registers the normal way 2716 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2717 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2718 } 2719 2720 // Kill projections on a branch should appear to occur on the 2721 // branch, not afterwards, so grab the masks from the projections 2722 // and process them. 2723 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2724 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2725 Node* use = n->fast_out(i); 2726 if (use->is_Proj()) { 2727 RegMask rm = use->out_RegMask();// Make local copy 2728 while( rm.is_NotEmpty() ) { 2729 OptoReg::Name kill = rm.find_first_elem(); 2730 rm.Remove(kill); 2731 anti_do_def( b, n, kill, false ); 2732 } 2733 } 2734 } 2735 } 2736 2737 // Check each register used by this instruction for a following DEF/KILL 2738 // that must occur afterward and requires an anti-dependence edge. 2739 for( uint j=0; j<n->req(); j++ ) { 2740 Node *def = n->in(j); 2741 if( def ) { 2742 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2743 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2744 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2745 } 2746 } 2747 // Do not allow defs of new derived values to float above GC 2748 // points unless the base is definitely available at the GC point. 2749 2750 Node *m = b->get_node(i); 2751 2752 // Add precedence edge from following safepoint to use of derived pointer 2753 if( last_safept_node != end_node && 2754 m != last_safept_node) { 2755 for (uint k = 1; k < m->req(); k++) { 2756 const Type *t = m->in(k)->bottom_type(); 2757 if( t->isa_oop_ptr() && 2758 t->is_ptr()->offset() != 0 ) { 2759 last_safept_node->add_prec( m ); 2760 break; 2761 } 2762 } 2763 } 2764 2765 if( n->jvms() ) { // Precedence edge from derived to safept 2766 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2767 if( b->get_node(last_safept) != last_safept_node ) { 2768 last_safept = b->find_node(last_safept_node); 2769 } 2770 for( uint j=last_safept; j > i; j-- ) { 2771 Node *mach = b->get_node(j); 2772 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2773 mach->add_prec( n ); 2774 } 2775 last_safept = i; 2776 last_safept_node = m; 2777 } 2778 } 2779 2780 if (fat_proj_seen) { 2781 // Garbage collect pinch nodes that were not consumed. 2782 // They are usually created by a fat kill MachProj for a call. 2783 garbage_collect_pinch_nodes(); 2784 } 2785 } 2786 2787 // Garbage collect pinch nodes for reuse by other blocks. 2788 // 2789 // The block scheduler's insertion of anti-dependence 2790 // edges creates many pinch nodes when the block contains 2791 // 2 or more Calls. A pinch node is used to prevent a 2792 // combinatorial explosion of edges. If a set of kills for a 2793 // register is anti-dependent on a set of uses (or defs), rather 2794 // than adding an edge in the graph between each pair of kill 2795 // and use (or def), a pinch is inserted between them: 2796 // 2797 // use1 use2 use3 2798 // \ | / 2799 // \ | / 2800 // pinch 2801 // / | \ 2802 // / | \ 2803 // kill1 kill2 kill3 2804 // 2805 // One pinch node is created per register killed when 2806 // the second call is encountered during a backwards pass 2807 // over the block. Most of these pinch nodes are never 2808 // wired into the graph because the register is never 2809 // used or def'ed in the block. 2810 // 2811 void Scheduling::garbage_collect_pinch_nodes() { 2812 #ifndef PRODUCT 2813 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2814 #endif 2815 int trace_cnt = 0; 2816 for (uint k = 0; k < _reg_node.Size(); k++) { 2817 Node* pinch = _reg_node[k]; 2818 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2819 // no predecence input edges 2820 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2821 cleanup_pinch(pinch); 2822 _pinch_free_list.push(pinch); 2823 _reg_node.map(k, NULL); 2824 #ifndef PRODUCT 2825 if (_cfg->C->trace_opto_output()) { 2826 trace_cnt++; 2827 if (trace_cnt > 40) { 2828 tty->print("\n"); 2829 trace_cnt = 0; 2830 } 2831 tty->print(" %d", pinch->_idx); 2832 } 2833 #endif 2834 } 2835 } 2836 #ifndef PRODUCT 2837 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2838 #endif 2839 } 2840 2841 // Clean up a pinch node for reuse. 2842 void Scheduling::cleanup_pinch( Node *pinch ) { 2843 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2844 2845 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2846 Node* use = pinch->last_out(i); 2847 uint uses_found = 0; 2848 for (uint j = use->req(); j < use->len(); j++) { 2849 if (use->in(j) == pinch) { 2850 use->rm_prec(j); 2851 uses_found++; 2852 } 2853 } 2854 assert(uses_found > 0, "must be a precedence edge"); 2855 i -= uses_found; // we deleted 1 or more copies of this edge 2856 } 2857 // May have a later_def entry 2858 pinch->set_req(0, NULL); 2859 } 2860 2861 #ifndef PRODUCT 2862 2863 void Scheduling::dump_available() const { 2864 tty->print("#Availist "); 2865 for (uint i = 0; i < _available.size(); i++) 2866 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2867 tty->cr(); 2868 } 2869 2870 // Print Scheduling Statistics 2871 void Scheduling::print_statistics() { 2872 // Print the size added by nops for bundling 2873 tty->print("Nops added %d bytes to total of %d bytes", 2874 _total_nop_size, _total_method_size); 2875 if (_total_method_size > 0) 2876 tty->print(", for %.2f%%", 2877 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2878 tty->print("\n"); 2879 2880 // Print the number of branch shadows filled 2881 if (Pipeline::_branch_has_delay_slot) { 2882 tty->print("Of %d branches, %d had unconditional delay slots filled", 2883 _total_branches, _total_unconditional_delays); 2884 if (_total_branches > 0) 2885 tty->print(", for %.2f%%", 2886 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2887 tty->print("\n"); 2888 } 2889 2890 uint total_instructions = 0, total_bundles = 0; 2891 2892 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2893 uint bundle_count = _total_instructions_per_bundle[i]; 2894 total_instructions += bundle_count * i; 2895 total_bundles += bundle_count; 2896 } 2897 2898 if (total_bundles > 0) 2899 tty->print("Average ILP (excluding nops) is %.2f\n", 2900 ((double)total_instructions) / ((double)total_bundles)); 2901 } 2902 #endif